NIMO: A Software Platform for Closed-Loop Materials Exploration with Diverse AI Algorithms

Abstract: Self-driving laboratories (SDLs), where artificial intelligence proposes subsequent experiments and robotic systems execute them, are rapidly becoming the vanguard of materials discovery. A critical bottleneck, however, lies in seamlessly bridging diverse AI algorithms tailored for specific exploration goals with the heterogeneous robotic hardware found across different laboratories. Here, we present NIMO, an open-source software platform designed to dissolve this barrier through three core paradigms: a modular AI-robot decoupling mediated via simple CSV file exchange, a discrete candidate-pool architecture that seamlessly absorbs domain knowledge, and a unified Python interface pre-loaded with twelve distinct AI algorithms. In this Perspective, we review the operational principles of each algorithm alongside six diverse SDL implementations driven by NIMO, covering electrolyte discovery, organic synthesis, thin-film exploration, fuel-cell process informatics, coffee-ring phase exploration, and legacy liquid-handling automation. One of these also demonstrates NIMO's seamless interoperability with the IvoryOS orchestration framework. To democratize autonomous science, we also introduce a no-code desktop application that enables intuitive, human-in-the-loop exploration for non-programmers. NIMO is freely available at https://github.com/NIMS-DA/nimo, offering a versatile, plug-and-play foundation to accelerate autonomous materials exploration across diverse experimental landscapes.

• The paper introduces a modular platform that decouples AI planning and robotics using CSV-based candidate pools to enforce experimental constraints.
• It validates twelve built-in AI algorithms across six self-driving lab deployments, demonstrating robust, scalable closed-loop experimentation.
• The architecture enhances hardware interoperability and democratizes autonomous materials discovery through open-source integration and no-code interfaces.

NIMO: Architecture and Foundations for Closed-Loop Materials Exploration

Motivation and System Architecture

The accelerating complexity in materials discovery, driven by the demand for novel compounds across industrial and scientific domains, increasingly necessitates autonomous experimental methodologies. Self-driving laboratories (SDLs), which integrate AI planning and robotic execution into closed experimental loops, are rapidly redefining the paradigm. However, the integration bottleneck between diverse AI algorithms—each suited for specific experimental goals—and heterogeneous laboratory hardware remains unresolved. "NIMO: A Software Platform for Closed-Loop Materials Exploration with Diverse AI Algorithms" (2606.15522) introduces an open-source, modular platform designed to dissolve these boundaries through three principal innovations: modular decoupling of AI and robotics via CSV file exchange, candidate-pool-driven search spaces that embed domain constraints as data, and a unified Python API with twelve built-in AI algorithms.

For maximum hardware and software interoperability, NIMO relies exclusively on CSV-based file exchange. This pragmatic approach decouples the AI planning module from robotic hardware, ensuring compatibility across languages (Python, LabVIEW, Visual Basic) and eliminating dependencies on in-memory APIs. NIMO's architecture thus readily supports integration with orchestration systems such as IvoryOS, which further broadens interface modalities through web-based drag-and-drop workflow design.

Candidate Pool and Data Abstraction

A defining aspect of NIMO is the candidate-pool abstraction. Unlike standard black-box optimization (BBO) paradigms, which operate within continuous bounded search boxes, NIMO inverts the paradigm by requiring users to pre-enumerate all experimentally feasible conditions in a single tabular dataset. The AI algorithm is strictly constrained to select candidates from this pool, never proposing infeasible experiments. This structure confers three key advantages: direct integration of domain knowledge and experimental constraints, seamless support for mixed variable types (continuous, discrete, categorical), and operational robustness in fully autonomous workflows.

Candidates are represented as rows in a CSV file, with columns covering experimental parameters and objective values. Completed experiments are annotated, while remaining candidates form the active pool for AI selection, enabling traceable closed-loop iterations.

Algorithm Portfolio and Technical Properties

NIMO exposes twelve AI algorithms through a uniform interface. These are categorized by exploration objectives:

• Bayesian Optimization: PHYSBO, a GP-based optimizer architected for discrete candidate pools, supports multiple acquisition functions (EI, PI, TS, HVPI, EHVI) and leverages parallelization and random feature mapping to scale to large search spaces.
• Process-Constrained Optimization: BOMP efficiently handles batch-level constraints where certain process variables are fixed across samples, significantly improving sample efficiency for manufacturing contexts.
• Threshold and Diversity-Based Sampling: NTS implements nested Thompson sampling to identify candidates above dynamic thresholds, tunable for aggression versus diversity.
• Combinatorial Exploration: COMBI targets composition-spread experiments via sequential optimization and gradient-based candidate selection, tailored for thin-film material synthesis.
• Target-Range Optimization: PTR proposes candidates maximizing the probability of falling within user-specified property windows, powerful for simultaneous multi-property constraints.
• Objective-Free Active Learning: BLOX maximizes property-space coverage via Stein discrepancy, crucial for constructing anomaly-rich materials libraries.
• Categorical Active Learning: PDC employs graph-based semi-supervised models (label propagation/spreading) for phase diagram construction, using uncertainty sampling to refine classification of categorical phase boundaries.
• Ranking and Transfer Learning: RSVM leverages ordinal relationships, robust against noisy objective values and compatible with transfer learning from related domains.
• Initial Sampling: RE (random), DOE (greedy, distance-based, D-optimal, LHS), and ES (exhaustive search) seed inaugural closed-loop cycles without training data.
• LLM-Based Planning: LLMEP uses LLMs for natural-language driven candidate selection, supporting multi-modal objectives, including qualitative descriptors.

Switching between algorithms is achieved via a single nimo.selection call, offering seamless repurposing across experimental campaigns.

Experimental Validation and Practical Deployments

NIMO's modularity and candidate-pool architecture have been validated across six SDL platforms:

• NAREE: High-throughput electrochemical platform (microplate-based) for autonomous electrolyte discovery. NIMO efficiently navigated a combinatorial space of 4,368 formulations, uncovering optimal lithium-metal anode electrolytes.
• CHEMSPEED: Automated synthesis integrated with SFC for end-to-end chemical reaction optimization, achieving rapid convergence on Suzuki-Miyaura coupling yields despite discrete categorical variables.
• COMBAT: Combinatorial sputtering platform for thin-film devices, employing manual transfers in a human-in-the-loop design and optimizing quinary alloy compositions for next-generation spintronic performance.
• ROPES: Automated fuel-cell process manufacturing, leveraging NIMO to refine operational sequences on a pilot line, bypassing traditional trial-and-error.
• Coffee Ring SDL: Robotic liquid handling combined with image-based classification, mapping phase boundaries in surfactant concentration space and demonstrating IvoryOS compatibility.
• BioDot: Integration with legacy VB-controlled liquid dispensers, requiring only lightweight file-exchange wrappers, enabling autonomous closed loops without touching existing software stacks.

These deployments emphasize NIMO's versatility in handling mixed-variable spaces, categorical objectives, legacy hardware, and collaborative orchestration.

Democratization and Interface Expansion

NIMO Desktop (Windows, macOS) provides a no-code graphical interface for human-in-the-loop optimization, mapping directly onto the candidate-pool abstraction and offering access to nine of the built-in algorithms. Parameterization and data logging are streamlined, ensuring operational parity with the developer-focused Python API. The platform is further integrated into web-based orchestration environments (IvoryOS) and supports emergent protocols (Model Context Protocol) for block-based workflow design.

Implications, Future Directions, and Outlook

NIMO distinguishes itself primarily via candidate-pool-driven decoupling, built-in algorithm diversity, and broad hardware interoperability. The operational architecture optimally supports domain-knowledge infusion, robust constraint imposition, and seamless deployment across heterogeneous laboratory environments—including legacy automation.

Theoretically, NIMO's abstraction aligns with "AI for Science," prioritizing algorithmic designs that embed practical constraints and real-world combinatorics natively. The future trajectory includes:

• Algorithmic Expansion: Multi-fidelity Bayesian optimization to integrate simulation and experiment, LLM-augmented algorithms for richer prior injection and critique, and ongoing inclusion of specialized strategies for hybrid experimental-theoretical loops.
• Interface Innovation: Ubiquitous access via graphical, web-based, and AI-driven platforms, supporting both fully autonomous and collaborative SDLs.
• Networked SDLs: Candidate-pool architecture enables distributed, cross-laboratory optimization with minimal data sharing, laying the groundwork for globally connected campaigns where labs specialize in complementary objectives or measurement modalities.

Practical implications encompass cost-effective upgrading of legacy environments, enhanced operational reliability, and democratized adoption of advanced experimental optimization in both industrial and academic contexts.

Conclusion

NIMO presents a rigorously modular, candidate-pool-centric platform for closed-loop materials exploration, delivering algorithmic breadth and hardware universality by design (2606.15522). Empirical results across six SDL deployments confirm performance scalability and versatility. The platform's data-driven abstraction—integrating domain constraints, heterogeneous objectives, and mixed hardware—sets the foundation for scalable, collaborative, and democratized autonomous experimentation. Future work is poised to extend algorithmic sophistication, interface models, and distributed SDL interconnectivity, aligning with the evolving needs of autonomous scientific discovery.